Text / Chat

GPT-5 Environmental Impact

OpenAI's next-gen flagship model

MODERATEPEER-REVIEWED
ArchitectureMultimodal Transformer (next-gen)Context256,000 tokensProviderOpenAI
18.4 WhEnergy / query
8.0 gCO₂ / query
66 mLWater / query
61x more thanvs Google search

Energy per query

18.4 Wh

61x more than a Google search (0.3 Wh)

CO2 per query

8.0 g

US East (Virginia) grid (450 gCO₂/kWh)

Water per query

66 mL

~15 queries to fill 1 litre

Processing location

Azure / Multi-cloud

Provider

OpenAI

Category

Text / Chat

Grid carbon intensity

450 g CO2/kWh (25% renewable)

How does GPT-5 compare?

Detailed Breakdown

Energy Consumption

GPT-5's per-query energy is genuinely uncertain, and the two best sources disagree by more than an order of magnitude. Jegham et al. (v6, 2025) measured ~18.35 Wh for a medium (~1000-token) response — a figure driven largely by GPT-5's default routing of many prompts through hidden reasoning steps. We report that benchmark measurement here as an upper-bound point estimate for a medium prompt. In contrast, the Joule 2026 production study (Oviedo et al.) analysed real fleet telemetry and found medians far lower — around 0.31 Wh per prompt, rising to ~3.91 Wh once test-time (reasoning) scaling is included — and concluded that benchmark-derived figures can overstate real production-fleet energy by roughly 4-20x. A typical GPT-5 prompt is therefore likely well below 18 Wh; we keep the peer-reviewed benchmark number as a stated-prompt-class upper bound while flagging this uncertainty prominently.

Power Source & Carbon

Runs across Azure, AWS, and Google Cloud — OpenAI's first truly multi-cloud model. This allows some region optimisation for carbon efficiency.

Water Usage

At ~66 mL per query on the Jegham benchmark basis — a figure that reflects the extended reasoning routing this medium-prompt measurement captures. A typical production prompt likely uses far less water (see the energy note on production-fleet vs benchmark estimates).

About GPT-5

GPT-5 is a text and chat model from OpenAI, released in August 7, 2025. OpenAI's next-gen flagship model. Each query uses 18.4 Wh of energy and produces 8.0 g of CO₂. That's 61x the energy of a Google search — reflecting the computational demands of text and chat.

GPT-5 in Context

100%
potential savings

The efficiency alternative

Gemini Nano performs the same type of task using just 0.01 Wh per query — 100% less energy than GPT-5. For a user sending 25 queries per day, switching would save 167.3 kWh per year.

1835 MWh
estimated daily

At global scale

With an estimated 10M+ daily users averaging 10 queries each, GPT-5 consumes roughly 1835 MWh of electricity per day — enough to power 61167 homes.

167.4 kWh
per year

Your yearly GPT-5 footprint

At 25 queries per day, your annual GPT-5 usage consumes 167.4 kWh — a meaningful fraction of household electricity. That produces 73.0 kg of CO₂.

Key Insights

Uses 5x more energy than the category average — reasoning models are inherently compute-intensive

What does your GPT-5 usage cost the planet?

Use our calculator to estimate your personal environmental footprint based on how often you use GPT-5.

Calculate My Compute

Frequently Asked Questions

How much energy does GPT-5 use per query?

Each GPT-5 query consumes approximately 18.4 Wh of energy. This is 61x more than a traditional Google search (~0.3 Wh).

What is GPT-5's carbon footprint?

Based on the carbon intensity of Azure / Multi-cloud, each query produces approximately 8.0 g of CO2. The grid in this region has a carbon intensity of 450 g CO2/kWh with 25% renewable energy.

How much water does GPT-5 use?

Each query consumes approximately 66 mL of water, primarily used for cooling the data centers that process the request.

How does GPT-5 compare to a Google search?

A GPT-5 query uses 61x more than a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while GPT-5 uses 18.4 Wh.

Technical Details

Architecture

Multimodal Transformer (next-gen)

Context window

256,000 tokens

Release date

2025-08-07

Open source

No

Training data cutoff

2025-06